[News] China’s AI Model Wave: Near-OpenAI Performance at Just 1% of the Cost?

With China’s AI companies on the rise, the global AI industry has entered a new Warring States era. As noted by ijiwei, following the emergence of DeepSeek, Alibaba’s Qwen platform has become as a top competitor to OpenAI, and it claims to match DeepSeek’s performance with minimal data.

As Chinese firms navigate trade restrictions and develop more efficient AI models, a wave of noteworthy innovations is gaining attention, including ByteDance’s Doubao AI model and Tencent’s Youdao AI chatbot.

A recent example is Baidu, which unveiled Ernie X1 and Ernie 4.5 just days ago. What’s most striking is that these models not only challenge OpenAI’s ChatGPT but also outprice China’s DeepSeek.

AI Model Showdown: Chinese Models Could Be 20-40x Cheaper

Prior to Baidu’s Ernie, DeepSeek has already stunned the market with the release of DeepSeek-V3 and DeepSeek-R1. And the company does not plan to stop there. According to Reuters, DeepSeek is fast-tracking the launch of R1’s successor. Initially set for early May, DeepSeek now reportedly aims to release R2 even sooner.

Reuters suggests that DeepSeek’s pricing is 20 to 40 times cheaper than equivalent models from OpenAI.

Now Baidu’s Ernie seems to offer competitive pricing as well. According to Business Insider, Ernie X1, a reasoning model, matches DeepSeek R1’s performance at half the cost, while Ernie 4.5, its latest foundation model and native multimodal model, claims to outperform GPT-4.5 in multiple benchmarks—at just 1% of the price.

Business Insider explains that tokens are the smallest data unit an AI model processes, and pricing depends on input and output token volume.

For Ernie 4.5, Baidu sets input token costs at 0.004 yuan per 1,000 and output tokens at 0.016 yuan per 1,000, as per Business Insider. By converting these prices to USD for comparison, while Baidu’s pricing undercuts OpenAI’s GPT-4.5, DeepSeek V3 is still slightly cheaper than Ernie 4.5.

In reasoning models, USD conversions show Ernie X1 as the cheapest, costing under 2% of OpenAI’s o1, according to Business Insider.

Future for China’s AI

While Baidu’s latest releases highlight intensifying U.S.-China AI competition and China’s increasing shift toward open-source models, U.S. tech giants still rely on heavy computing power to train models, leading to higher costs for developers.

Another South China Morning Post report notes that OpenAI’s o1 charges $60 per million output tokens—nearly 30 times the cost of DeepSeek-R1.

Furthermore, on March 20th, OpenAI introduced o1-pro, a pricier upgrade via its API platform. Using more compute for better responses, it’s OpenAI’s most expensive model yet. According to Techcrunch, OpenAI charges $150 per million input tokens (~750,000 words) and $600 per million output tokens—twice the cost of GPT-4.5 for input and 10 times that of standard o1.

In addition to price advantage, China’s AI labs seem to be closing their technology gaps with Western counterparts. As highlighted by ijiwei, when OpenAI launched o1 in December, 2024, a comparable model, DeepSeek R1, has been developed within months.

TrendForce notes that China’s AI market is expected to develop in two key directions in light of ongoing U.S. chip export restrictions. First, AI-related companies will accelerate investments in domestic AI chips and supply chains. Large Chinese CSPs, for instance, will continue procuring available H20 chips while also ramping up the development of proprietary ASICs for deployment in their data centers.

Second, China will leverage its existing internet infrastructure to compensate for hardware limitations with software-based solutions. DeepSeek exemplifies this approach by breaking from conventional methods and adopting model distillation technology to enhance AI applications.

Read more

• [News] 20 Global Tech Giants Announce Integration with DeepSeek
• AI Server Growth in 2025 Faces Uncertainties, DeepSeek Effect to Boost AI Inference Share, Says TrendForce

(Photo credit: Baidu)